Evaluation of 3D markerless pose estimation accuracy using openpose and depth information from a single RGB-D camera

Abstract

Safe and efficient Human-Robot Collaboration (HRC) requires recognizing human collaborators intention. Hand pose kinematics early on an ongoing movement can provide information for predicting future human actions. Accurate state-of-the-art methods used for human hand pose estimation are either marker-based or make use of multiple cameras set around the workspace. These approaches introduce inconvenience to the user, necessitate calibration and are bounded to the specific set-up and workspace. On the other hand, using a single RGB-D camera would be less obtrusive for the user and less cumbersome to install. In this work, we use OpenPose to extract 2D keypoints from the RGB raw image and we combine them with the depth information acquired from the RGB-D camera to obtain 3D hand poses. We evaluate the accuracy and discrimination ability of our method in ten different static poses.